Learning Power Grid Outages With Higher-Order Topological Neural Networks

Yuzhou Chen, Roshni Anna Jacob, Yulia R. Gel, Jie Zhang, H. Vincent PoorIEEE Transactions on Power Systems 2023 · DOI

TL;DR

This paper introduces HOT-Nets, a novel deep learning model that uses topological data analysis (persistent homology) and higher-order graph structures (simplicial complexes) to significantly improve the detection of power grid outages. By capturing complex network interdependencies beyond simple node-to-node connections, HOT-Nets achieves state-of-the-art performance in classifying outage scenarios, enhancing power grid resilience.

What Problem Does This Paper Address?

The increasing frequency of extreme weather events and cyber-physical threats has made the resilience of power distribution networks (DNs) a critical concern. Detecting outages in these networks is crucial for timely restoration and preventing cascading failures. However, DNs are complex, vast, and often only partially observable due to the high cost of sensor deployment. Traditional methods struggle with this complexity and sparsity. Existing machine learning approaches often overlook the inherent topological structure of the network, which is vital for understanding how component failures propagate and affect the system. This paper aims to bridge this gap by developing a model that effectively leverages both topological features and higher-order network interactions for robust outage detection.

Methods

The core innovation of this work is the development of Higher-Order Topological Neural Networks (HOT-Nets). This model integrates two key concepts:

  • Persistent Homology (PH): This technique from topological data analysis is used to extract "topological signatures" of the distribution network. PH identifies and quantifies shape features (like connected components, holes, etc.) that are invariant to continuous deformations, providing essential descriptors of the network's structure. Different "filtrations" (e.g., based on node degree, betweenness, or closeness) are used to capture multi-scale topological properties, generating "persistence images" (PIs).
  • Simplicial Neural Networks (SNNs) with Hodge-Laplacian Analytics: Instead of traditional graph convolutions that only consider direct connections (1-hop neighbors), HOT-Nets extends convolutions to "simplicial complexes." These complexes represent higher-order structures (like triangles or tetrahedrons) within the network. The Hodge-Laplacian operator is used to efficiently process these higher-order structures, allowing the model to capture complex, multi-node interactions that are crucial in power grids.

The HOT-Nets architecture first uses PH to generate topological features (PIs) and then employs a Convolutional Neural Network (CNN) to process these PIs. Simultaneously, a higher-order simplices convolution (HoSC) module processes the network's simplicial structure. An attention mechanism is then used to adaptively weigh the importance of these different topological representations (from PH and SNNs) before feeding them into a final classifier.

Datasets

The HOT-Nets model was validated on three standard distribution network test cases:

  • IEEE 37-bus feeder
  • IEEE 123-bus feeder
  • 342-bus low voltage network

Two distinct case studies were used for evaluation:

  1. Full Observability: Sensors are assumed to be placed at all buses, providing complete network state information.
  2. Partial Observability: Sensors are placed at only 45-50% of the buses and lines, simulating a more realistic, sparse sensing scenario.

The datasets were generated synthetically by simulating line failures and load profile variations using the OpenDSS simulation software. This approach allows for controlled generation of various outage scenarios.

Key Insights and Ideas

  • Topology Matters, Especially Higher-Order Topology: The paper strongly emphasizes that understanding the complex, multi-node interactions (beyond simple pairwise connections) is critical for accurate outage detection in power grids. Traditional graph methods often fall short because they don't capture these higher-order relationships.
  • Persistent Homology as a Powerful Feature Extractor: Using persistent homology to generate "topological signatures" provides a robust way to represent the essential structural characteristics of the network, even with sparse sensor data.
  • Simplicial Complexes Capture Network Complexity: Extending graph convolutions to simplicial complexes allows the model to learn from more complex substructures within the network, leading to a richer representation of system behavior.
  • Attention Mechanism for Adaptive Learning: The attention mechanism is key to effectively integrating information from different topological representations (from PH and SNNs), allowing the model to focus on the most relevant features for outage detection.
  • Bridging Data-Driven and Physics-Informed Approaches: While primarily data-driven, the use of topological features derived from the network structure implicitly incorporates some understanding of the underlying physics and connectivity.

Key Results

  • Substantial Performance Gains: HOT-Nets consistently outperformed 9 state-of-the-art methods across all tested networks and scenarios.
  • Relative Gains up to 14.04%: The paper reports significant relative gains in classification accuracy, with the highest observed in the case studies.
  • Robustness to Sparsity: The model demonstrated strong performance even under partial observability, highlighting its effectiveness in real-world, sensor-limited scenarios.
  • Ablation Studies Confirm Component Importance: Experiments showed that each component of HOT-Nets (Hodge Laplacian, multifiltrations for PIs, and attention mechanism) contributes significantly to its overall performance.

Limitations and Future Work

  • Synthetic Data: While robust, the current evaluation relies on synthetically generated data. Future work could involve validation on real-world outage data.
  • Outage Location and Impact Estimation: The current model focuses on outage classification (detecting if an outage exists). Future research aims to extend this to determining the location of outages and estimating the scale of disruption (e.g., energy not served).
  • Generalization to Other Systems: While applied to power grids, the HOT-Nets framework could potentially be adapted for resilience analysis in other complex cyber-physical systems.
  • Optimal Design: The authors suggest exploring the use of HOT-Nets for optimal design problems within cyber-physical systems.

Summary updated

Abstract

With the increase in cyber-physical threats and extreme weather events, resilience of the power system has become a problem of utmost societal importance. In this paper, we propose a novel approach for resilience improvement of power distribution networks, based on the notions of persistent homology and simplicial neural networks (SNNs) which are new directions in graph learning. In particular, tools of persistent homology allow us to capture the most essential topological descriptors of the distribution network. In turn, extending the convolutional operation to simplicial complexes on the distribution network, using the Hodge-Laplacian analytics, enables us to describe complex interactions among multi-node higher order graph substructures. Such higher order graph substructures are of particular importance in distribution networks, since a change in power demand at a load bus (or the power supplied from a substation) will produce a corresponding perturbation in nodal variables (such as the bus voltages) and edge variables (such as branch currents). We validate our new Higher-Order Topological Neural Networks (HOT-Nets) model for contingency classification of three test distribution networks, the IEEE 37-bus feeder, IEEE 123-bus feeder, and the 342-bus low voltage network. Our experiment results on two case studies (i.e., (i) with sensors placed at all the buses alone in the networks and (ii) with partial observability in the networks) indicate that HOT-Nets substantially outperforms 9 state-of-the-art methods, yielding relative gains of up to 14.04% in terms of system resilience classification.
BibTeX
@article{chen2023learning,
  author = {Yuzhou Chen and Roshni Anna Jacob and Yulia R. Gel and Jie Zhang and H. Vincent Poor},
  title = {{Learning Power Grid Outages With Higher-Order Topological Neural Networks}},
  journal = {IEEE Transactions on Power Systems},
  year = {2023},
  doi = {10.1109/tpwrs.2023.3266956},
}
Figure